laya-nli-conflict-v10-l2 — research archive (round 10, v9-config rerun), NOT delivered

⚠️ Research archive — NOT a delivered model. This checkpoint failed its round's acceptance gates and was never shipped. The current production head is slow-stack/laya-nli-memory-conflict (v4). Uploaded 2026-10-02 for provenance/backup while round 11 (multi-run verdict protocol) waits for Kaggle GPU quota.

Round 10 (2026-10-01) of the laya NLI memory-conflict head program by modusensus. L2 = an independent rerun of the v9 corpus and configuration (kernel-direct, complete artifact set), completing the three-run same-config evidence together with the v9 round run and the S2 baseline (laya-nli-conflict-v10-s2).

Headline results

• main val 0.889 (✗), new-10 9/10 (✗), bias diagnosis 12/14 (✗) ⇒ 9/12 gates
• Why this run matters: across the three byte-identical-config runs (v9 / S2 / L2) — main val 0.9050 / 0.8940 / 0.8890 (1.6pp swing), τ(noul) 1.03–1.19, and B2 = 0.889 / 0.3877 / 0.6951 — a 50pp range crossing the 0.5 decision line. Same config, both "pass" and "fail" ⇒ single-run B2 readings have no decision power; round 11 therefore preregisters a multi-run median protocol with a 9-case isomorphic family instrument.

Artifacts

file	value
model.safetensors	SHA256 e557d46b…b98d15 (full hash in archive_sha256_manifest.txt)
rl_agent_config.json	τ(noul) = 1.0284; encoder jhu-clsp/mmBERT-base; bf16
metrics.json	val_accuracy 0.889, val_ece 0.0408, n_val 1000, no_rl true
val_probs.json	frozen-val probability dump (calibration analyses)

Provenance

• Training: Kaggle GPU kernel daphnelaurent/laya-nli-conflict-ce (three-version sequence), dataset daphnelaurent/nli-conflict-pairs v15
• Round record & full gate table: kaggle_eval/HANDOFF_NLI_V10.md